In recent years, facing the problems of carbon emissions and environmental protection, new energy such as PV power generation has attracted more and more attention. Moreover, global dependence on the use of electrical energy is increasing rapidly, which also increases people’s interest in developing renewable energy [
1]. Among a series of renewable energy sources, such as solar, wind, hydroelectric and geothermal energy, wind, hydroelectric and geothermal energy are mostly limited by regions, seasons and climates, and their utilization and production rates are low. On the contrary, solar energy systems play a vital role in providing stable power demand due to their advantages of being environmentally friendly, low-carbon energy, with safe operation, noiseless impact and low cost [
2]. The International Renewable Energy Agency reported that the global solar PV net capacity additions from 2019 to 2021 were 109.6 GW, 134.9 GW and 151 GW, respectively [
3]. In PV systems, various anomalies usually lead to losses of electricity production and affect the working efficiency and even the operational safety of modules. These anomalies may be caused by harsh weather in actual use, or mechanical damage during manufacturing, installation and transportation. These damages will also shorten the actual service life of PV modules [
4]. Therefore, it is still a problem that needs to be solved to effectively monitor the health of PV cells and ensure their performance and safety [
5].
Several methods, such as EC (Electrical Characterization) [
6], EL (Electroluminescence) imaging [
7] and infrared imaging [
8] methods have been adopted to detect defects in PV cells. The principle of EC is to find out the faults of PV cells by analyzing the I-V characteristics, but some small faults will hardly affect the characteristics, and this method cannot locate the fault area. In contrast, the methods of EL and infrared imaging technology more easily recognize and locate faults. By obtaining the gray image of the tested object in the dark, EL imaging can effectively identify the existing micro-crack fault. However, due to the single image color, fault identification is time-consuming and expensive, so EL imaging is only suitable for small-scale fault detection. The infrared imaging method uses different colors to represent the temperature distribution of the tested object, so as to judge the existence form of faults. Compared with EL imaging, infrared imaging is more suitable for large-scale fault detection.
For large-scale applications, several infrared thermography (IRT) methods linked to passive [
9] and active [
10] approaches have been applied to detect various faults in PV cells. He et al. [
11] used electromagnetic induction infrared thermography (EIIT) to detect scratches, hot spots and other faults in PV cells, and analyzed the thermography sequence through Fast Fourier Transform (FFT), Independent Component Analysis (ICA) and Principal Component Analysis (PCA). Breitenstein et al. [
12] used dark lock-in thermography (DLIT) technology to analyze the leakage current phenomenon, local efficiency and I-V characteristics of PV cells. In particular, he used DLIT to evaluate short circuit current density imaging, and experimental results showed that this method can improve the accuracy of DLIT-based local efficiency analysis of solid cells [
13]. In another study, the “Local I-V” method for evaluating DLIT images was applied for high efficiency monocrystalline silicon solar cells, and the results indicate that DLIT can be used to image and quantify the local inhomogeneous dark current contribution of PV cells [
14]. Straube et al. [
15] used illuminated lock-in thermography (ILIT) to analyze the shunt phenomenon in PV cells. The efficiency of DLIT and ILIT is compared by Frühauf et al. [
16]. With the increasing demand for large-scale production quantities of PV cells, fault detection needs to change from traditional visual detection to automatic detection. In particular, the application of AI (artificial intelligence) algorithms and deep learning makes fault detection and classification of PV cells’ thermography images more efficient. Akram et al. [
17] proposed an isolated-learning-model-based automatic defects detection method for PV cells using infrared images. A light CNN was designed to train the isolated learning model from scratch with an accuracy of 98.67%. When developing the model transfer deep learning method, a base model was first pre-trained using an EL image dataset of PV cells. The next step was fine-tuned training on an infrared image dataset, which achieved an accuracy of 99.23%. However, the time cost of this model takes 1 h and 17 min, and the research focuses on the datasets of two types of faults. Chen et al. [
18] proposed a visual fault detection method with a multi-spectral deep convolutional neural network (CNN) to analyze the light spectrum features of PV cell color images. The method can detect existing surface faults with an accuracy of 90%. However, the proposed multi-spectral CNN has weak feature extraction capabilities for small defects such as cracks. Deitsch et al. [
19] used CNN and SVM to detect PV cell EL images. The results indicate that the CNN model has a higher accuracy than SVM, with an average accuracy of 88.42%. However, the trained CNN model requires higher computational power and consumes a large amount of resources. Tang et al. [
20] used an image generation method and a CNN-model-based PV cell faults classification using electroluminescence images for the samples. The image generation method combines traditional image processing and Generative Adversarial Network (GAN) characteristics. The accuracy of the proposed CNN model was about 80% for micro-crack, fault-free and break-fault types, respectively. However, this study mainly inspected the same type of defects, such as micro-cracks and finger interruption and breaks. Akram et al. [
21] used a light convolutional neural network model based identify fault in EL images of PV cells, and various data augmentation methods were evaluated to solve data scarcity. The model obtained an ideal detection result of 93.02%, and it took only 8.07 ms to predict one image. Cipriani et al. [
22] proposed a novel method based on CNN to realize the classification of dust and hotspot faults in PV systems through thermography technology. The experimental results achieved an accuracy of 98%. However, the dataset used in this study was too sparse. Wang et al. [
23] proposed a hybrid algorithm by combining the symmetrized dot pattern (SDP) with a convolutional neural network (CNN) for PV module fault recognition. Three faults such as poor welding, breakage and bypass diode were discussed. The experimental results show that the proposed algorithm can capture the fault signals effectively, display them in images and recognize the PV modules’ fault types accurately. The literature review indicates that although deep learning has shown some performance in EL studies, existing studies still have limitations such as high computational cost, low performance and detection of specific defects, etc. In order to better overcome these limitations and ensure the healthy performance of PV cells, it is possible to achieve efficient and low-cost PV fault detection in infrared images by adopting more suitable CNN models and generalization strategies. Therefore, we integrated residual structural units in the series network model and propose a CNN model based on infrared image features of PV cells to achieve automatic classification of cell faults and predict their power generation efficiency and potential safety issues.
In summary, the current research mainly analyzes and obtains the existing faults through the EL images of PV cells in outdoor service, while there is less research on the detection and classification of faults that may occur during the manufacturing process of cells. This process requires faster detection efficiency and fault classification speed. A new CNN model is proposed to process the cells’ infrared image dataset to meet this requirement. First of all, infrared images of PV cells with faults were collected, and offline data augmentation technology was adopted to solve the problem of data scarcity and improve the generalization ability of the network. Secondly, the designed CNN model integrated the residual units, which can fully extract the deep features of PV cells. Finally, the proposed model was verified on the pre-treated infrared PV cells dataset, and the results show that the proposed model can better perform the task of PV cell fault classification.